/ops-social-planner — engine-agnostic planned-content viewer skill

/ops-social-planner — engine-agnostic planned-content viewer

Load ops-rules before acting. Public repo (no personal data). Outbound: one draft → one approval → one send. If AskUserQuestion / Workflow are missing, follow Rule 10 in ops-rules (Hermes: numbered options / two-turn Telegram card; delegate_task).

Read-only dashboard of every scheduled post and ad, grouped identity/project → channel → time, regardless of posting engine. Personal/founder identity and project brands stay strictly separated (see /ops-socials identity rules). See SPEC.md for the full design.

What it does

1. bin/ops-social-planner collect reads $PREFS_PATH/preferences.json (marketing.social_identities.personal.* + marketing.projects.*.social), dispatches a per-engine fetcher keyed on social.engine.primary, normalizes to one schema, and writes $OPS_DATA_DIR/social-planner/state.json (owner data — never committed, Rule 0).
2. It then serves the static UI (ui/) on localhost and opens the browser.

Engines

engine.primary	Source	Status
typefully	GET /v2/social-sets/{id}/drafts (Bearer; key from ~/.config/typefully/config.json)	✅ wired
upload-post	GET /api/uploadposts/schedule (Apikey; key from engine.upload_post.api_key_ref)	✅ wired
meta-graph / meta-ads	Graph Ads API per project BM/token	🔌 hook (UI shows "pending")
google-ads	GAQL per project customer id	🔌 hook
null / unprovisioned	—	shown fail-closed (0 items)

Adding an engine = adding one fetch<Engine>() in bin/ops-social-planner + a dispatch branch. No UI change.

Run it

"${CLAUDE_PLUGIN_ROOT}/bin/ops-social-planner"                 # collect → serve → open
"${CLAUDE_PLUGIN_ROOT}/bin/ops-social-planner" collect          # regenerate state.json only
"${CLAUDE_PLUGIN_ROOT}/bin/ops-social-planner" serve --port 7937  # serve existing state

The UI renders ui/state.sample.json (synthetic, PII-free) when no live state exists, so it works with zero credentials — useful for the PR preview.

Agent enrichment path (optional, when invoked as a skill)

The headless collector covers posts on wired engines. When richer output is wanted, the agent can:

• Pull ads via MCP (mcp__meta__*, Google Ads) per project and merge kind:"ad" items into state.json.
• Replace the heuristic rationale with an LLM-written one (slot + sequence + channel intent).
• Re-run serve to refresh. Do not publish or edit from here — routing/mutation belongs to /ops-socials + /ops-marketing.

Guarantees

• Read-only. No secrets in the client. Server binds 127.0.0.1 only.
• Identity separation honored: personal vs project never merged; unprovisioned projects render fail-closed.
• No owner data committed — only ui/ + bin/ + SPEC.md + the synthetic fixture.